docs: add validated DGX evidence cards
Browse files- .gitattributes +2 -0
- README.md +4 -0
- assets/dgx-qwen38-evidence.png +3 -0
- assets/dgx-qwen38-performance.png +3 -0
.gitattributes
CHANGED
|
@@ -1,3 +1,5 @@
|
|
| 1 |
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 2 |
model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
|
| 3 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
| 1 |
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
| 2 |
model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
|
| 3 |
tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
| 4 |
+
assets/dgx-qwen38-performance.png filter=lfs diff=lfs merge=lfs -text
|
| 5 |
+
assets/dgx-qwen38-evidence.png filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -31,6 +31,8 @@ Validated on one NVIDIA DGX Spark with TensorFold:
|
|
| 31 |
|
| 32 |
These are workload-specific validation results, not a universal model speed or capability claim. Full protocol and limitations are below.
|
| 33 |
|
|
|
|
|
|
|
| 34 |
## Provenance
|
| 35 |
|
| 36 |
| Field | Value |
|
|
@@ -60,6 +62,8 @@ The script explicitly skips:
|
|
| 60 |
|
| 61 |
No quantizer is invoked. All 21 output shard sizes match the pinned base checkpoint. Sampled byte hashes matched the base for routed NVFP4 expert payloads, CSF fixed/exception scales and outer scales; sampled target attention tensors differed as expected. This was a sampled integrity check, not an exhaustive hash of all 373,248 routed-expert keys.
|
| 62 |
|
|
|
|
|
|
|
| 63 |
## Performance validation
|
| 64 |
|
| 65 |
### System and runtime
|
|
|
|
| 31 |
|
| 32 |
These are workload-specific validation results, not a universal model speed or capability claim. Full protocol and limitations are below.
|
| 33 |
|
| 34 |
+

|
| 35 |
+
|
| 36 |
## Provenance
|
| 37 |
|
| 38 |
| Field | Value |
|
|
|
|
| 62 |
|
| 63 |
No quantizer is invoked. All 21 output shard sizes match the pinned base checkpoint. Sampled byte hashes matched the base for routed NVFP4 expert payloads, CSF fixed/exception scales and outer scales; sampled target attention tensors differed as expected. This was a sampled integrity check, not an exhaustive hash of all 373,248 routed-expert keys.
|
| 64 |
|
| 65 |
+

|
| 66 |
+
|
| 67 |
## Performance validation
|
| 68 |
|
| 69 |
### System and runtime
|
assets/dgx-qwen38-evidence.png
ADDED
|
Git LFS Details
|
assets/dgx-qwen38-performance.png
ADDED
|
Git LFS Details
|